{"id":"876e0219-2fb2-4411-a3b4-c34e48aa92df","arxiv_id":"2505.04611","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"m-PGibbs, a marginalized particle Gibbs algorithm, replaces the Gibbs parameter update with a conditional sequential Monte Carlo step and shows more robust mixing than PMMH for small particle numbers.","lead":"This paper introduces a new version of particle Gibbs that removes a slow update step by averaging over the parameter with a conditional sequential Monte Carlo kernel. On a simple state-space example, the method explores the parameter space more reliably than PMMH when the number of particles is small.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's final-index accept/reject (lines 13–14) does not sample K ~ Cat(W_T), so the CSMC kernel as printed is not obviously π-invariant; Algorithm 2's validity and Figure 1 rest on this step.","rationale":"The reader identifies the invariance of the non-Markovian CSMC kernel as the weakest assumption. I agree that this is the load-bearing point, but the manuscript's problem is sharper than a missing proof: the pseudocode for the final ancestor selection in Algorithm 1 appears inconsistent with the standard CSMC construction it cites. The central claim that Algorithm 2 is a valid MCMC kernel depends directly on the CSMC update being exactly π-invariant; if lines 13–14 are taken literally, the output index distribution is not Cat(W_T) and the standard detailed-balance argument fails. The empirical results in Figure 1 cannot repair this because a biased sampler's acceptance rate is not an estimate of anything at the target distribution. The issue is concrete and likely fixable as a typo or pseudocode error, so I would not move from the reader's CONDITIONAL verdict to REJECT; however, the required revision is now specific: correct the final-index sampling, prove the resulting kernel is invariant (or cite the standard result), and re-verify the empirical claims. My agreement with the reader is partial because my concern is a concrete defect in the stated algorithm rather than merely an absence of proof for a generally valid construction.","tokens_in":12289,"tokens_out":24221,"duration_ms":240177,"concrete_test":"Check π-invariance of Algorithm 1 on the smallest nontrivial case: N=2, T=1, with an arbitrary known target. Enumerate the kernel K(x'|x) implied by lines 3–16 with B=0 and verify ∫K(x'|x)π(x)dx=π(x'). A simpler sufficient diagnostic: for a fixed final weight vector W_T, compute the marginal probability of output index 1 under lines 13–14; with W_T=(0.4,0.6) the printed procedure gives P(K=1)=1/3 instead of W_T^1=0.4. If this is a typo, replace line 14 with 'accept with probability 1-W_T^1' and rerun the Section 4 experiment to confirm that the reported gains are not an artifact of the uncorrected kernel.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 2 defines Algorithm 1 as a π-invariant CSMC kernel, and Algorithm 2 as a collapsed Gibbs sampler inherits correctness from this. In the standard CSMC/PGibbs construction, the final output index K must be drawn from the normalized final weights W_T, i.e. P(K=j)=W_T^j. Lines 13–14 instead sample k'_T from the weights with particle 1 removed and accept it with probability W_T^1/W_T^{k'_T}, retaining particle 1 on rejection. For N=2, W_T=(0.4,0.6), this procedure gives P(K=1)=1-0.4/0.6=1/3 and P(K=2)=2/3, not P(K=1)=0.4 and P(K=2)=0.6. The acceptance probability needed to match Cat(W_T) is 1-W_T^1, independent of k'_T. Unless the printed line is a typo, the final ancestor distribution is wrong, so the standard invariance argument does not apply and the m-PGibbs chain in Algorithm 2 targets the wrong joint distribution; the 'valid MCMC kernel' claim and the empirical acceptance-rate comparison in Section 4 are therefore unsupported for the algorithm as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces 'marginalized particle Gibbs' (m-PGibbs), an MCMC method for joint parameter and trajectory inference in Feynman–Kac/state-space models. The construction augments the parameter with an auxiliary variable u and M parameter particles θ^{1:M}, one of which (index l) is tied to the trajectory in the target (6). The paper defines a non-Markovian target for the trajectory obtained by marginalizing over l (Section 3.2), proposes to sample it with a conditional SMC kernel (Algorithm 1), and then resamples l from its posterior (Algorithm 2). It claims m-PGibbs is a valid collapsed Gibbs sampler, that the parameter update is asymptotically a Barker acceptance step, and that the method mixes better than PMMH for small particle numbers. Section 4 gives an acceptance-rate comparison on a Gaussian state-space model with code provided.","tokens_in":12522,"tokens_out":16535,"duration_ms":151743,"significance":"If the invariance claims were established, m-PGibbs would be a useful addition to the particle MCMC toolbox: it avoids the trajectory/parameter Gibbs coupling of PGibbs, uses CSMC rather than PIMH as the underlying 'state-only' kernel, and has a transparent asymptotic interpretation in terms of Barker acceptance. The paper is clearly written, gives code, and honestly states the main technical caveats (no theoretical analysis, single empirical example). However, the correctness of the algorithm as written is compromised by an invalid final-index sampling step in Algorithm 1, and the non-Markovian invariance is asserted rather than proved; these issues currently block acceptance.","major_comments":[{"comment":"Lines 13–14 do not sample the final trajectory index from the normalized weights W_T. For N=2 and W_T=(0.4,0.6), the printed procedure draws k'_T=2 with probability one and accepts with probability 0.4/0.6, giving P(K=1)=1/3 instead of 0.4; in general P(K=1)=1-Σ_{j=2}^N [W_T^j/(1-W_T^1)](W_T^1/W_T^j)=1-(N-1)W_T^1/(1-W_T^1), which is not a valid probability when W_T^1 ≥ 1/N. The required acceptance probability for the conditional proposal over {2,...,N} is 1-W_T^1 (or K should be drawn directly from Cat(W_T)). Because Algorithm 2 in §3.3 and the empirical results in §4 inherit their correctness from Algorithm 1, the validity claims are unsupported as printed.","section":"§2, Algorithm 1, lines 13–14"},{"comment":"The central validity assertion—that Algorithm 2 defines a valid MCMC kernel—rests on the claim that Algorithm 1 is π-invariant for the non-Markovian marginalized target π_T(x_{0:T} | u, θ^{1:M}). No proof or explicit theorem is provided; the paper says in §3.3 that the incremental term (12) can be evaluated backwards in time, but (12) is a forward recursion for π_{t-1}(l | x_{0:t-1}, u, θ^{1:M}). The backward-sampling weights in Algorithm 1, lines 19–23, use γ_T(x_{0:t}^n, x'_{t+1:T})/γ_t(x_{0:t}^n); for this non-Markovian target the validity of that step requires an argument that the manuscript does not give. Please state and prove a lemma (or give a precise reference to a theorem covering exactly this non-Markovian CSMC construction) establishing invariance of the kernel in Algorithm 1 with the proposal and potential of Section 3.2, and clarify how π_T(l|x,u,θ^{1:M}) is obtained from the backward pass. This is load-bearing because the collapsed-Gibbs derivation of Algorithm 2 reduces to this invariance.","section":"§3.3 and §2"},{"comment":"The empirical claim that m-PGibbs explores the parameter space more efficiently than PMMH for smaller N is based on a single run per algorithm, raw acceptance rates, no error bars, and no comparison with the closest existing method (parameter-eliminated PGibbs of Wigren et al., 2019), even though Remark 2 connects m-PGibbs to that work. The text notes that m-PGibbs costs about twice as much as PMMH for the same N but the figure does not report acceptance rate per computation or effective sample size; the assertion that higher acceptance rate implies higher ESS (Tierney, 1998) is not generally valid when proposals and costs differ. Please provide repeated runs with Monte Carlo error, cost-adjusted measures (e.g., ESS per second or per iteration), and include the Wigren et al. benchmark or justify its omission.","section":"§4, Figure 1"}],"minor_comments":[{"comment":"There are typos in the abstract and introduction: 'algortihm', 'conversly', and '£rom' in Section 1.1.","section":"Abstract and Introduction"},{"comment":"The sentence 'Both three choices are valid' should read 'All three choices are valid.'","section":"§3.4"},{"comment":"The notation g_t(x_{0:t}; θ^{1:M}, u, l') is confusing because l' is the output index and the right-hand side of (11) does not depend on l'; please remove l' from the notation or explain its role.","section":"Eq. (11) and Algorithm 2 line 5"},{"comment":"The reference 'Pitt and and' is incomplete, and the entry 'Finke and Thiery (to appear, 2023)' should be formatted consistently.","section":"References"},{"comment":"The calibration value is written 'τ = 0.152'; if this means 0.15^2, please clarify the notation.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The main obstruction is the Algorithm 1 final-index error, which is a one-line fix (accept with probability 1-W_T^1 or sample K from Cat(W_T)). I recommend major revision rather than rejection because the rest of the construction appears coherent and the empirical claim could be repaired. I also note that per Remark 2 the novelty is partly a special case of Wigren et al. (2019); the contribution would be strengthened by an explicit, direct comparison and a clearer statement of what is new beyond the non-Markovian CSMC implementation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the idea is genuinely neat: write the PGibbs parameter update as sampling a categorical index, marginalize it out, and run CSMC on the resulting non-Markovian target. The Barker acceptance interpretation is a useful way to see why it behaves differently from PMMH. Second, the algorithm as printed appears to have a bug in the final selection step of CSMC, and that bug breaks the invariance claim the whole paper rests on.\n\nThe paper is honest about limited novelty: Remark 2 notes the construction is a special case of Wigren et al. (2019). What's new is the specific proposal mechanism and the packaging, which is fine. The exposition is clear, and the empirical section makes a real point: m-PGibbs is robust to small N where PMMH collapses. Code is provided.\n\nNow the soft spots, in proportion. The biggest issue is Algorithm 1, lines 13-14. The stated procedure samples k'_T from the weights excluding particle 1, then accepts with probability W^1_T / W^{k'_T}_T. That does not sample from Cat(W_T). For N=2 with W=(0.4,0.6), this gives P(K=1)=1/3, not 0.4. The correct acceptance probability to match Cat(W_T) is 1-W^1_T, independent of k'. So either the line is a typo or the CSMC kernel is not π-invariant. If it's a typo, it's easily fixed; if not, the collapsed Gibbs argument in Algorithm 2 is unsupported. The paper does not flag this step as approximate, so this is a load-bearing flaw.\n\nSecond, the backward sampling step for the marginalized target is asserted rather than derived. The target (8) is non-Markovian in x, and the simplification used for ordinary state-space models does not apply. The claim that π_T(l|...) can be computed as a byproduct of backward sampling needs a real derivation, not a hand-wave.\n\nThird, the empirical work is thin: one data set, no error bars, no comparison to Wigren et al., and the runtime difference between PMMH and m-PGibbs is acknowledged but not folded into the efficiency comparison.\n\nWho is this for? Particle MCMC people who want a parameter-elimination alternative to PMMH. The idea is worth taking seriously. I would send it to a serious referee, but the referee should be asked to verify the CSMC kernel and demand a proof or explicit derivation of the backward sampling step. As it stands, the paper is not ready to be used as is, but it's not a dud.","headline":"Neat idea, load-bearing flaw in Algorithm 1's index sampling; send to a careful referee.","tokens_in":13048,"tokens_out":6334,"would_cite":false,"duration_ms":57165,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M05","62F15","65C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces m-PGibbs, a collapsed particle-MCMC kernel that marginalizes the parameter update in particle Gibbs, and reports that it explores parameter space more efficiently than PMMH for small particle counts.","keywords":["particle Gibbs","marginalized particle Gibbs","conditional sequential Monte Carlo","state-space models","pseudo-marginal Metropolis-Hastings","Barker acceptance","collapsed Gibbs sampler","parameter inference"],"falsifier":"Run m-PGibbs on a small linear-Gaussian state-space model where the exact posterior is available from a Kalman smoother, using a proposal $q(u|\\theta),q(\\theta'|u)$ for which the Barker acceptance probability can be computed in closed form; if the empirical stationary distribution differs from the exact posterior, or the acceptance rate does not approach the theoretical Barker probability as $N$ grows, the invariance claim is refuted.","tokens_in":12041,"feed_emoji":"🎲","tokens_out":9749,"duration_ms":88851,"temperature":0.7,"pith_summary":"Sampling a state-space model's latent path and its parameters by alternating updates—the usual particle Gibbs scheme—can mix slowly when the path and parameters are strongly correlated, while particle marginal Metropolis-Hastings (PMMH) avoids the alternation but relies on noisy likelihood estimates and can stall when the particle count is small. The paper introduces m-PGibbs, a collapsed version of particle Gibbs: it introduces an auxiliary variable and several parameter copies, forms an augmented target in which a parameter index can be integrated out, and runs conditional sequential Monte Carlo directly on the marginalized trajectory target. The claim is that the resulting kernel targets the exact joint posterior of trajectory and parameter, and that in the paper's example it keeps accepting parameter moves at particle counts where PMMH's acceptance is essentially zero. If correct, this would give joint parameter-trajectory inference the stability advantages that conditional sequential Monte Carlo already has over particle independent Metropolis-Hastings in the fixed-parameter setting.","feed_headline":"Marginalized particle Gibbs outperforms PMMH at small particle counts","feed_subtitle":"By collapsing the Gibbs step, m-PGibbs keeps parameter chains moving when PMMH's acceptance collapses.","key_machinery":"The machinery is an auxiliary-variable construction that makes the parameter update collapse. The parameter proposal is split as $q(\\theta'|\\theta)=\\int q(\\theta'|u)q(u|\\theta)\\,du$; with $M$ parameter copies $\\theta^{1:M}$ and a flag $l$, the paper forms the joint distribution $\\pi(x_{0:T},\\theta^{1:M},u,l)\\propto\\pi_T(x_{0:T}|\\theta_l)q(u|\\theta_l)\\prod_{j\\neq l}q(\\theta_j|u)$. Conditional on $u$ and $\\theta^{1:M}$, the trajectory target is the mixture $\\pi_T(x_{0:T}|u,\\theta^{1:M})=\\sum_l\\pi_T(x_{0:T},l|u,\\theta^{1:M})$, and the running posterior over $l$ follows the one-step reweighting (12). A conditional sequential Monte Carlo kernel targeting this non-Markovian mixture is run, and $l$ is then sampled from its terminal posterior, with backward-sampling weights obtained by evaluating (12) backwards in time. The object doing the work is the categorical posterior over $l$, which makes the model computationally Markovian and keeps the per-iteration cost at $O(TNM)$.","core_discovery":"The central discovery is that the Gibbs sweep in particle Gibbs can be collapsed: instead of alternating a trajectory update and a parameter update, one can run a conditional sequential Monte Carlo kernel on a marginalized target and then sample the parameter index from a closed-form categorical distribution. The paper calls the resulting kernel m-PGibbs and claims it is a valid MCMC kernel for the joint target $\\pi_T(x_{0:T},\\theta)$, with the parameter effectively updated by a Barker acceptance step in the limit of infinitely many particles rather than by Metropolis-Hastings. Its empirical section shows a state-space example in which m-PGibbs sustains positive acceptance for small $N$ while PMMH does not, and only loses to PMMH for very large $N$, where PMMH approaches its ideal 27% acceptance and m-PGibbs approaches Barker's roughly 18%.","pith_inferences":["The same auxiliary-index collapse could in principle be embedded in online or recursive particle parameter samplers, not just in offline MCMC, whenever the parameter proposal admits the $u$-decomposition; the paper does not explore this.","Because the categorical posterior over $l$ is conjugate to any likelihood model, m-PGibbs extends parameter elimination beyond the conjugate settings where earlier work removed parameters exactly, at the price of carrying $M$ parameter copies.","Replacing the Barker acceptance by a Metropolis-Hastings step—which the paper notes as possible but does not develop—would combine m-PGibbs's small-$N$ robustness with PMMH's higher asymptotic acceptance ceiling.","For $M>2$ the method inherits the known limitations of multiproposal samplers in static regimes, so the paper's recommendation $M=2$ is conservative; better multi-try or gradient-informed choices for $q(u|\\theta)$ might change the trade-off."],"forward_implications":["m-PGibbs offers a parameter-MCMC update whose acceptance is far less sensitive to the number of particles $N$; in the paper's example it keeps exploring at $N$ values where PMMH's acceptance is near zero.","Because the limiting acceptance is Barker's rather than Metropolis-Hastings, m-PGibbs has a lower asymptotic acceptance ceiling than PMMH, so for very large $N$ a well-tuned PMMH can eventually become more statistically efficient.","The algorithm inherits the backward-sampling machinery of conditional sequential Monte Carlo, so the stability advantages of CSMC over PIMH for fixed parameters carry over to joint parameter-trajectory inference.","With the recommended $M=2$, the per-iteration cost is roughly twice that of PMMH, so the small-$N$ advantage is not an artifact of ignoring computational cost."],"supporting_citations":[{"why":"introduces conditional sequential Monte Carlo and particle Gibbs, whose theory and structure m-PGibbs extends.","marker":"Andrieu et al. (2010)"},{"why":"provides the pseudo-marginal interpretation of PMMH and the acceptance-rate comparison against ideal Metropolis that frames the empirical study.","marker":"Andrieu and Roberts (2009)"},{"why":"supplies convergence results for particle Gibbs used to justify the infinite-particle behaviour of the marginalized CSMC step.","marker":"Chopin and Singh (2015)"},{"why":"establishes uniform ergodicity and exact-simulation limits of conditional SMC kernels, invoked for the asymptotic recovery of the idealized sampler.","marker":"Andrieu et al. (2018)"},{"why":"gives stability results for conditional backward sampling that motivate using backward sampling in the CSMC step.","marker":"Lee et al. (2020)"},{"why":"shows how parameters can be eliminated in particle Gibbs; m-PGibbs is explicitly read as a special case via the categorical prior over $l$.","marker":"Wigren et al. (2019)"},{"why":"provides the collapsed Gibbs sampler construction that makes m-PGibbs a valid collapsed kernel.","marker":"Liu (1994)"},{"why":"defines the Barker acceptance step that m-PGibbs recovers in the infinite-particle limit and that sets its asymptotic acceptance ceiling.","marker":"Barker (1965)"},{"why":"supplies the multi-proposal auxiliary-variable construction used to split $q(\\theta'|\\theta)$ into $q(\\theta'|u)q(u|\\theta)$.","marker":"Tjelmeland (2004)"},{"why":"provides the benchmark state-space model, the data, and the calibrated proposal scale used in the empirical comparison.","marker":"Chopin and Papaspiliopoulos (2020)"}],"fun_headline_variants":["Collapsing the Gibbs step boosts particle MCMC","Marginalized PGibbs: faster mixing than PMMH","No Gibbs step, just Barker: new particle MCMC","Skipping the Gibbs sweep improves particle MCMC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the conditional sequential Monte Carlo step remains an exact, invariant update for the marginalized target that depends on the whole past trajectory, and that evaluating the one-step weight update (12) backwards gives the correct backward-sampling weights; the paper asserts this without proof, and standard theory does not cover such non-Markovian targets.","fun_headline_variants_meta":{"raw":{"variants":["Collapsing the Gibbs step boosts particle MCMC","Marginalized PGibbs: faster mixing than PMMH","No Gibbs step, just Barker: new particle MCMC","Skipping the Gibbs sweep improves particle MCMC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00039,"raw_usage":{"total_tokens":2053,"prompt_tokens":947,"completion_tokens":1106,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":1041}},"tokens_in":563,"tokens_out":1106,"duration_ms":8664,"temperature":1.0,"reasoning_tokens":1041,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:24:07.728905+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run m-PGibbs on a small linear-Gaussian state-space model where the exact posterior is available from a Kalman smoother, using a proposal $q(u|\\theta),q(\\theta'|u)$ for which the Barker acceptance probability can be computed in closed form; if the empirical stationary distribution differs from the exact posterior, or the acceptance rate does not approach the theoretical Barker probability as $N$ grows, the invariance claim is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces conditional sequential Monte Carlo and particle Gibbs, whose theory and structure m-PGibbs extends."},{"cited_title":"and Roberts, G","cited_arxiv_id":null,"evidence_quote":"provides the pseudo-marginal interpretation of PMMH and the acceptance-rate comparison against ideal Metropolis that frames the empirical study."},{"cited_title":"and Singh, S","cited_arxiv_id":null,"evidence_quote":"supplies convergence results for particle Gibbs used to justify the infinite-particle behaviour of the marginalized CSMC step."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"establishes uniform ergodicity and exact-simulation limits of conditional SMC kernels, invoked for the asymptotic recovery of the idealized sampler."},{"cited_title":"S., and Vihola, M","cited_arxiv_id":null,"evidence_quote":"gives stability results for conditional backward sampling that motivate using backward sampling in the CSMC step."},{"cited_title":"S., Murray, L., and Lindsten, F","cited_arxiv_id":null,"evidence_quote":"shows how parameters can be eliminated in particle Gibbs; m-PGibbs is explicitly read as a special case via the categorical prior over $l$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the collapsed Gibbs sampler construction that makes m-PGibbs a valid collapsed kernel."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the Barker acceptance step that m-PGibbs recovers in the infinite-particle limit and that sets its asymptotic acceptance ceiling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the multi-proposal auxiliary-variable construction used to split $q(\\theta'|\\theta)$ into $q(\\theta'|u)q(u|\\theta)$."},{"cited_title":"and Papaspiliopoulos, O","cited_arxiv_id":null,"evidence_quote":"provides the benchmark state-space model, the data, and the calibrated proposal scale used in the empirical comparison."}],"review_version":1}